Dice Question Streamline Icon: https://streamlinehq.com

Competitiveness of observation-only AI-DOP systems versus IFS at medium range

Determine whether an end-to-end data-driven forecast system trained and initialized exclusively from Earth System observations (i.e., an AI-DOP system such as GraphDOP) can achieve medium-range forecast skill competitive with the ECMWF Integrated Forecast System (IFS), a state-of-the-art physics-based model.

Information Square Streamline Icon: https://streamlinehq.com

Background

The paper introduces GraphDOP, an AI-DOP model trained and initialized solely from observations without using physics-based reanalysis inputs. While GraphDOP produces skilful forecasts up to five days and shows competitive performance in some regions and variables, overall medium-range skill does not yet match the IFS. The authors explicitly note the unresolved question of whether observation-only, end-to-end systems can ultimately compete with IFS at the medium range.

References

While the results presented herein give good cause for optimism, it remains to be determined whether an end-to-end data-driven system trained and initialized exclusively from observations can compete at the medium range with a state-of-the-art physics-based system such as the IFS.

— GraphDOP: Towards skilful data-driven medium-range weather forecasts learnt and initialised directly from observations (2412.15687 - Alexe et al., 20 Dec 2024) in Section 7, Discussion and outlook